Pith. sign in

REVIEW 3 major objections 4 minor 27 references

VIBE: Video-Input Brain Encoder for fMRI Response Modeling

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read VIBE, a two-stage multimodal Transformer, predicts fMRI responses to movies with a mean parcel-wise correlation of 0.3225 on held-out Friends and 0.2125 on six out-of-distribution films, beating the official ridge baseline by roughly 0.12.

desk verdict VIBE is a solid, credible Algonauts-winning system with a large margin over the ridge baseline; the post-challenge score tweaks are small enough not to threaten the main conclusion. read the letter →

arxiv 2507.17958 v2 pith:SIUHL66C submitted 2025-07-23 cs.LG cs.AIcs.CV

Daniel Carlström Schad , Shrey Dixit , Janis Keck , Viktor Studenyak , Aleksandr Shpilevoi , Andrej Bicanski This is my paper
classification cs.LGcs.AIcs.CV
keywords fMRIencodingmultimodaltransformernaturalisticstimulivideo-to-brainpredictivecodingCNeuroModfeatureattributionensemblelearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

VIBE is a two-stage Transformer that predicts whole-brain fMRI activity from movies by fusing features extracted from audio, video, and transcripts. Trained on 65 hours of the CNeuroMod movie dataset and averaged over 20 seeds, it reaches mean parcel-wise Pearson correlations of 0.3225 on the held-out Friends season and 0.2125 on six out-of-distribution films, versus official ridge regression baselines of 0.2033 and 0.0895. The paper argues that the combination of domain-specific open-source feature extractors, temporal transformers with rotary embeddings, network-specific models, and ensembling is what makes these gains, and that the model generalizes across animation, black-and-white footage, and foreign languages. The result matters because it shows naturalistic brain activity can be predicted from the multimodal content of experience, and it produces parcel-level maps of which features drive which brain networks.

What carries the argument

The load-bearing components are two Transformer stages. The Modality Fusion Transformer performs per-TR cross-attention over projected features from Qwen2.5-14B text, BEATs and Whisper audio, SlowFast and V-JEPA 2 video, and Qwen2.5-Omni, each reduced to 256 dimensions and combined with a 256-dimensional subject embedding. The Prediction Transformer then applies two layers with Rotary Positional Embeddings (RoPE, relative position via rotations) to model temporal structure across 1.49-second TRs. Ensembling across 20 seeds and training separate visual/default-mode networks on the same architecture are the main post-hoc levers; removing the causal mask gives a small additional gain.

What would settle it

Shift the feature-to-fMRI alignment by one TR (1.49 s) on a held-out movie and re-score the ensemble; a sharp drop in correlation would confirm the assumed timing is load-bearing, while little change would suggest the model mostly exploits static content correlations.

Watch

Extended reading notes

Core claim

The paper's central discovery is that movie-driven fMRI responses can be predicted with a pipeline that separates modality fusion from temporal decoding: a single-layer modality fusion transformer merges features at each TR, and a two-layer prediction transformer with rotary positional embeddings maps the fused sequence to BOLD activity in 1000 cortical parcels. The best configuration uses 20-seed ensembling, dedicated models for visual and default-mode parcels, and no causal mask, so the model may attend to future time points. It reports mean Pearson r = 0.3225 on Friends S07 and r = 0.2125 on six out-of-distribution films, exceeding the official baseline by about 0.12 in both settings. The paper also shows that text features dominate language and default-mode parcels, audio features peak in auditory cortex, and visual features cover occipital and motion-sensitive areas, while a single omni-model feature set adds little once specialized features are present.

Load-bearing premise

The central claim depends on the supplied stimulus-to-fMRI timing being correct, since the model is trusted to learn the brain's delayed response by itself; if true delays vary across regions or subjects, the reported correlations would be misestimated.

Editorial extensions

If this is right

  • The reported scores (0.3225 in-distribution, 0.2125 out-of-distribution) set a concrete benchmark for encoding models on the CNeuroMod movies.
  • Long-context transcript features improved default-mode predictions, suggesting narrative-level structure is encoded in those parcels.
  • Allowing future context helped slightly, supporting predictive-coding accounts of naturalistic viewing, while the small effect size qualifies how much the brain anticipates.
  • Dedicated visual and default-mode models outperform a single whole-cortex model, indicating that parcel grouping by functional network is a useful inductive bias.
  • The model transfers to animated, black-and-white, and French-language films despite training only on English Friends episodes, implying content-agnostic representations.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • One could test whether the same pipeline transfers to other naturalistic modalities, such as first-person video or audiobooks, by swapping the front-end feature extractors while keeping the two-stage transformer fixed.
  • Because omni-model features added little once specialized encoders were present, a single strong multimodal encoder might suffice if its internal layers are probed the way the paper probes V-JEPA and Qwen.
  • The parcel-level Shapley maps could be turned into explicit predictions: for example, lesions or altered states that change default-mode engagement should change text-feature contributions more than auditory-feature contributions.
  • If the temporal-alignment assumption holds, VIBE's lack of an explicit HRF suggests that transformer-based encoders can replace hemodynamic convolution in future encoding models.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper presents VIBE, a two-stage Transformer architecture for predicting fMRI BOLD responses from naturalistic movie stimuli. It fuses features from open-source pretrained models across text (Qwen2.5 14B), audio (BEATs, Whisper-V3), video (V-JEPA 2, SlowFast), and an omni model (Qwen2.5 Omni), using a modality-fusion transformer followed by a prediction transformer with rotary positional embeddings. Trained on 65 hours of CNeuroMod data from the Algonauts 2025 challenge, ensembled over 20 seeds, VIBE obtains mean parcel-wise Pearson correlations of 0.3225 on the in-distribution Friends S07 split and 0.2125 on six out-of-distribution films, compared to official baseline ridge correlations of 0.2033 and 0.0895. The paper also reports that an earlier version achieved 0.3198 and 0.2096, winning Phase 1 and placing second overall in the challenge. Additional analyses include functional-network-specific models, removal of the causal mask in the prediction transformer, and Multiperturbation Shapley-Value Analysis (MSA) of feature contributions across brain parcels.

Significance. If the reported results hold, VIBE represents a substantial advance in fMRI encoding for naturalistic video, with large gains over the challenge baseline and plausible, anatomically consistent feature-attribution maps. The paper's strengths include the use of openly available feature extractors, a publicly released code repository with configuration files, a clear comparison to an independent baseline, and external validation through the leaderboard outcomes. The MSA attribution is a useful addition that generates testable hypotheses about modality-specific cortical responses. However, the manuscript lacks statistical uncertainty quantification for its central claims, and the post-challenge bug-fixed scores are not anchored to a documented, audited pipeline, so the precise headline numbers should be treated with caution.

major comments (3)
  1. [Section 1 and Section 4] The central performance figures (0.3225 in-distribution, 0.2125 out-of-distribution) are post-challenge results that differ from the official leaderboard scores (0.3198 and 0.2096). The manuscript states that 'minor implementation bugs' were fixed and that 'the most effective Phase-2 techniques' were back-ported into the Phase-1 pipeline, but it gives no details on what the bugs were, how the fixes were selected, or whether the test set was consulted during this selection. This matters because if the fixes were chosen based on leaderboard feedback, the reported improvements could be a selection artifact. The official, verifiable leaderboard scores already demonstrate a large margin over baseline, so the qualitative conclusion survives, but the manuscript should either describe the bug fixes and their selection process in sufficient detail to rule out test-set information leakage, or report the official scores as the primary results and relegate the post-challenge numbers to a clearly separated supplementary claim.
  2. [Section 3.2.2 and Section 4.1] The claim that removing the causal mask improves performance by approximately 0.002 is not accompanied by any measure of variability or statistical test. With a 20-seed ensemble, it would be straightforward to report the distribution of single-seed differences or a bootstrap confidence interval on the ensemble difference. Without such quantification, the improvement is difficult to distinguish from noise, and the subsequent interpretation about 'backwards causality' and predictive coding (Section 5) is not supported at the reported magnitude. This issue also affects the other attribution claims in Section 4.1 (ensembling gain of 0.011–0.012, network-specific models gain of 0.003–0.004), which are stated as point estimates without error bars.
  3. [Section 3.2.3] The decision to model the hemodynamic response implicitly, rather than via explicit HRF convolution or a learnable 1-d convolution, is based on the observation that the explicit approaches 'both decreased performance.' However, no quantitative results or significance levels are provided for this comparison, despite it being a load-bearing architectural choice. If the implicit temporal attention can only capture a fixed lag, and the true BOLD delay varies across parcels or subjects, the reported correlations could be systematically misestimated. The authors should either provide validation numbers for the alternative HRF treatments or temper the claim that the implicit approach is definitively superior.
minor comments (4)
  1. [Section 2] The list of films in the training set is oddly punctuated ('Life, Hidden Figures, and The Bourne Supremacy' with 'Life' appearing twice as a film and as a word), and the sentence should be rephrased for clarity.
  2. [Section 4.1] The text says 'removing the causal mask from the prediction transformer yielded an additional 0.002 percentage points' but the metric is Pearson r (dimensionless), not percentage points. This should be corrected to avoid confusion.
  3. [Figure 2] Figure 2 shows validation scores for network-specific models but the text does not provide the actual numerical values or specify which validation split was used; adding the numbers or a table would make the comparison more transparent.
  4. [Section 3.1.1] The use of a manually created one-line description for each clip, including the movie name, duration, and occasionally the director's name, is described in a single sentence. If this text is part of the model input, it should be explicitly stated whether the description was included in the Qwen input for all clips or only some, and whether the manual descriptions were created before or after seeing the test data, since this could affect the interpretation of the text-feature contributions.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: VIBE's headline correlations are held-out measurements with externally pretrained features and an independent organizer baseline; the only self-citation is a non-load-bearing post-hoc attribution tool.

full rationale

The central claim — mean parcel-wise Pearson r of 0.3225 on Friends S07 and 0.2125 on six out-of-distribution films versus organizer baselines of 0.2033 and 0.0895 — is a held-out performance measurement against external test data, not a quantity derived from the model's own inputs. All features come from open-source models (Qwen2.5, BEATs, Whisper, SlowFast, V-JEPA, LaBSE) pretrained on non-fMRI datasets, and the target parcels come from the CNeuroMod/Algonauts pipeline, so no input is defined in terms of the output or vice versa. The loss (L = LPearson + 0.03·LMSE) aims at correlation on training blocks while the reported numbers are computed on unseen blocks, so no fitted parameter is renamed as a prediction. Architecture choices (RoPE, 20-seed ensembling, network-specific models, unmasked attention) were selected on a held-out validation season and then reported as post-hoc ablations; this is standard model development, not circularity. The manuscript's only self-citation is the MSA toolkit (refs [21] and [22], co-author Shrey Dixit), used in Section 4.2 for post-hoc parcel-wise attribution; it does not affect prediction scores and is therefore not load-bearing. The post-challenge 'bug fix' passage in the Introduction after the abstract ('we revisited our codebase, identified minor implementation bugs...') reports improved numbers (0.3225/0.2125) alongside the official leaderboard values (0.3198/0.2096) without stating how the fixes were selected; this is a reproducibility and verification caveat about possible test-set-informed selection, not a circular reduction of the kind defined here, and it does not change the qualitative margin over the independent baseline. Similarly, the absence of an explicit HRF (Section 3.2.3) is a modeling-assumption risk rather than a circular step. Score 1 reflects the single minor non-load-bearing self-citation; the derivation chain itself is self-contained against an external benchmark.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper adds no new physical or biological entities. The central claim rests on domain assumptions about the dataset, the fixed feature extractors, the implicit HRF handling, and the validity of removing the causal mask. Hyperparameters such as the loss weight and context windows are hand-chosen and not swept systematically.

free parameters (5)
  • loss weight lambda_MSE = 0.03
    Hand-set weight for the MSE term in the total loss L = LPearson + lambda_MSE * LMSE; no reported sweep.
  • V-JEPA context window = 6 s (3 s before to 1.51 s after the TR)
    Hand-selected window for visual feature extraction, stated in Section 3.1.2.
  • audio context window = 10 s with 1.49 s stride
    Hand-selected to match BEATs and Whisper training clip lengths, stated in Sections 3.1.3 and 3.1.4.
  • number of fusion and prediction layers = 1 fusion, 2 prediction
    Architecture depth chosen by validation experimentation; no ablation table is provided.
  • network-specific ensemble selection = Visual, Default, and All models
    After experimentation, three parcel-masked models were selected and combined; this is a modeling choice rather than a fitted scalar.
assumptions (5)
  • domain assumption fMRI BOLD responses are linearly predictable from stimulus features, and Pearson correlation is the appropriate evaluation metric.
    Inherited from the Algonauts challenge setup; used throughout Sections 2 and 4.
  • domain assumption The provided preprocessing (MNI152 space, Schaefer 1000-parcel parcellation, subject 4 exclusion) is correct.
    Used without re-analysis; described in Section 2.
  • domain assumption Pretrained feature extractors (Qwen2.5, BEATs, Whisper, SlowFast, V-JEPA, Qwen2.5-Omni, LaBSE) provide useful representations for brain prediction.
    Assumed throughout Section 3.1; no experiments test whether alternative extractors would work.
  • ad hoc to paper The hemodynamic response can be learned implicitly by the prediction transformer rather than modelled explicitly.
    The authors found that HRF convolution and a learnable 1-d convolution both decreased performance (Section 3.2.3), so the model relies on its internal temporal attention.
  • ad hoc to paper Allowing the prediction transformer to attend to future time points does not introduce test-set leakage.
    The causal mask was removed and produced a 0.002 gain, which the authors interpret as predictive coding (Section 3.2.2), but they do not rule out other explanations or check for leakage through the manually added clip descriptions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VIBE: Video-Input Brain Encoder for fMRI Response Modeling." pith.science (2026). https://pith.science/paper/SIUHL66C

@misc{pith2026250717958,
  author       = {Pith},
  title        = {Pith review of: VIBE: Video-Input Brain Encoder for fMRI Response Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SIUHL66C}},
  note         = {Machine review of arXiv:2507.17958}
}
read the original abstract

We present VIBE, a two-stage Transformer that fuses multi-modal video, audio, and text features to predict fMRI activity. Representations from open-source models (Qwen2.5, BEATs, Whisper, SlowFast, V-JEPA) are merged by a modality-fusion transformer and temporally decoded by a prediction transformer with rotary embeddings. Trained on 65 hours of movie data from the CNeuroMod dataset and ensembled across 20 seeds, VIBE attains mean parcel-wise Pearson correlations of 0.3225 on in-distribution Friends S07 and 0.2125 on six out-of-distribution films. An earlier iteration of the same architecture obtained 0.3198 and 0.2096, respectively, winning Phase-1 and placing second overall in the Algonauts 2025 Challenge.

Figures

Figures reproduced from arXiv: 2507.17958 by the authors.

Figure 1
Figure 1. Overview of the VIBE architecture. (1) We extract a diverse set of visual features from different layers of [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Validation scores for models separately trained on subsets of parcels corresponding to separate functional [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Model scores for in and out of distribution test sets. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Multiperturbation Shapley-Value Analysis (MSA) of parcel-wise feature contributions. Warm colors mark parcels where a feature set improves prediction accuracy the most. Text features from Qwen 2.5 14B dominate higher-order language and default-mode areas as well as par…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 12 canonical work pages

  1. [1]

    Daniel L. K. Yamins and James J. DiCarlo. Using goal-driven deep learning models to understand sensory cortex. Nature Neuroscience, 19(3):356–365, March 2016. ISSN 1546-1726. doi:10.1038/nn.4244. URL https://www.nature.com/articles/nn.4244. Publisher: Nature Publishing Group

    doi:10.1038/nn.4244 2016

  2. [2]

    Encoding model of temporal processing in human visual cortex

    Anthony Stigliani, Brianna Jeska, and Kalanit Grill-Spector. Encoding model of temporal processing in human visual cortex. Proceedings of the National Academy of Sciences , 114(51):E11047–E11056, December

    work page

  3. [3]

    Marcel A. J. van Gerven. A primer on encoding models in sensory neuroscience. Journal of Mathematical Psychology, 76:172–183, February 2017. ISSN 0022-2496. doi:10.1016/j.jmp.2016.06.009. URL https: //www.sciencedirect.com/science/article/pii/S0022249616300487

    work page doi:10.1016/j.jmp.2016.06.009 2017

  4. [4]

    Gifford, Domenic Bersch, Marie St-Laurent, Basile Pinsard, Julie Boyle, Lune Bellec, Aude Oliva, Gemma Roig, and Radoslaw M

    Alessandro T. Gifford, Domenic Bersch, Marie St-Laurent, Basile Pinsard, Julie Boyle, Lune Bellec, Aude Oliva, Gemma Roig, and Radoslaw M. Cichy. The Algonauts Project 2025 Challenge: How the Human Brain Makes Sense of Multimodal Movies, January 2025. URL http://arxiv.org/abs/2501.00504. arXiv:2501.00504 [q-bio]

    arXiv 2025

  5. [5]

    Boyle, Boris Pinsard, and others

    John A. Boyle, Boris Pinsard, and others. The Courtois project on neuronal modelling - 2020 data release, June 2020. Published: Poster presented at the 2020 Annual Meeting of the Organization for Human Brain Mapping (OHBM)

    work page 2020

  6. [6]

    Johnsrude, and Adrian M

    Matthew Brett, Ingrid S. Johnsrude, and Adrian M. Owen. The problem of functional localization in the human brain. Nature Reviews Neuroscience, 3(3):243–249, March 2002. ISSN 1471-0048. doi:10.1038/nrn756. URL https://doi.org/10.1038/nrn756

    work page doi:10.1038/nrn756 2002

  7. [7]

    Local-Global Parcellation of the Human Cerebral Cortex from Intrinsic Functional Connectivity MRI

    Alexander Schaefer, Ru Kong, Evan M Gordon, Timothy O Laumann, Xi-Nian Zuo, Avram J Holmes, Simon B Eickhoff, and B T Thomas Yeo. Local-Global Parcellation of the Human Cerebral Cortex from Intrinsic Functional Connectivity MRI. Cerebral Cortex, 28(9):3095–3114, July 2017. ISSN 1047-3211. doi:10.1093/cercor/bhx179. URL https://doi.org/10.1093/cercor/bhx179

    doi:10.1093/cercor/bhx179 2017

  8. [8]

    Qwen2.5 Technical Report, January 2025

    Qwen, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Ti...

    arXiv 2025

Show all 27 references
  1. [9]

    V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning, 2025

    Mido Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Mojtaba, Komeili, Matthew Muckley, Ammar Rizvi, Claire Roberts, Koustuv Sinha, Artem Zholus, Sergio Arnaud, Abha Gejji, Ada Martin, Francois Robert Hogan, Daniel Dugas, Piotr Bojanowski, Vasil Khalidov, Pat...

    2025 arXiv

  2. [10]

    BEATs: Audio Pre-Training with Acoustic Tokenizers

    Sanyuan Chen, Yu Wu, Chengyi Wang, Shujie Liu, Daniel Tompkins, Zhuo Chen, Wanxiang Che, Xiangzhan Yu, and Furu Wei. BEATs: Audio Pre-Training with Acoustic Tokenizers. In Proceedings of the 40th International Conference on Machine Learning , pages 5178–5193. PMLR, July 2023. ...

    2023

  3. [11]

    Robust Speech Recognition via Large-Scale Weak Supervision, December 2022

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust Speech Recognition via Large-Scale Weak Supervision, December 2022. URL http://arxiv.org/abs/ 2212.04356. arXiv:2212.04356 [eess]

    2022 arXiv

  4. [12]

    Ariel Goldstein, Haocheng Wang, Leonard Niekerken, Mariano Schain, Zaid Zada, Bobbi Aubrey, Tom Sheffer, Samuel A. Nastase, Harshvardhan Gazula, Aditi Singh, Aditi Rao, Gina Choe, Catherine Kim, Werner Doyle, Daniel Friedman, Sasha Devore, Patricia Dugan, Avinatan Hassidim, Mi...

    2025

  5. [13]

    SlowFast Networks for Video Recognition, October 2019

    Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast Networks for Video Recognition, October 2019. URL http://arxiv.org/abs/1812.03982. arXiv:1812.03982 [cs]

    2019 arXiv

  6. [14]

    Qwen2.5-Omni Technical Report, March

    Jin Xu, Zhifang Guo, Jinzheng He, Hangrui Hu, Ting He, Shuai Bai, Keqin Chen, Jialin Wang, Yang Fan, Kai Dang, Bin Zhang, Xiong Wang, Yunfei Chu, and Junyang Lin. Qwen2.5-Omni Technical Report, March

  7. [15]

    Language-agnostic BERT Sentence Embedding, 2020

    Fangxiaoyu Feng, Yinfei Yang, Daniel Cer, Naveen Arivazhagan, and Wei Wang. Language-agnostic BERT Sentence Embedding, 2020. URL https://arxiv.org/abs/2007.01852. Version Number: 2

    2020 arXiv

  8. [16]

    RoFormer: Enhanced Transformer with Rotary Position Embedding, November 2023

    Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. RoFormer: Enhanced Transformer with Rotary Position Embedding, November 2023. URL http://arxiv.org/abs/2104. 09864. arXiv:2104.09864 [cs]

    2023 arXiv

  9. [17]

    Rajesh P. N. Rao and Dana H. Ballard. Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2(1):79–87, January 1999. ISSN 1546-1726. doi:10.1038/4580. URL https://www.nature.com/articles/nn019...

    1999 doi

  10. [18]

    Predictive coding under the free-energy principle

    Karl Friston and Stefan Kiebel. Predictive coding under the free-energy principle. Philosophical Trans- actions of the Royal Society B: Biological Sciences , 364(1521):1211–1221, May 2009. ISSN 0962-8436. doi:10.1098/rstb.2008.0300. URL https://www.ncbi.nlm.nih.gov/pmc/article...

    2009

  11. [19]

    Lindquist, Ji Meng Loh, Lauren Y

    Martin A. Lindquist, Ji Meng Loh, Lauren Y . Atlas, and Tor D. Wager. Modeling the Hemodynamic Response Function in fMRI: Efficiency, Bias and Mis-modeling. Neuroimage, 45(1 Suppl):S187–S198, March 2009. ISSN 1053-8119. doi:10.1016/j.neuroimage.2008.10.065. URL https://www.ncb...

    2009 doi

  12. [20]

    B. T. Thomas Yeo, Fenna M. Krienen, Jorge Sepulcre, Mert R. Sabuncu, Danial Lashkari, Marisa Hollinshead, Joshua L. Roffman, Jordan W. Smoller, Lilla Zöllei, Jonathan R. Polimeni, Bruce Fischl, Hesheng Liu, and Randy L. Buckner. The organization of the human cerebral cortex es...

    2011

  13. [21]

    Kording, and Claus C

    Shrey Dixit, Kayson Fakhar, Fatemeh Hadaeghi, Patrick Mineault, Konrad P. Kording, and Claus C. Hilgetag. Who Does What in Deep Learning? Multidimensional Game-Theoretic Attribution of Function of Neural Units, June 2025. URL http://arxiv.org/abs/2506.19732. arXiv:2506.19732 [cs]

    2025 arXiv

  14. [22]

    MSA: A compact Python package for Multiperturbation Shapley value Analysis., 2021

    Kayson Fakhar and Shrey Dixit. MSA: A compact Python package for Multiperturbation Shapley value Analysis., 2021. URL https://github.com/kuffmode/msa. Publication Title: GitHub repository

    2021

  15. [23]

    A value for n-person games

    Lloyd S Shapley and others. A value for n-person games. 1953. Publisher: Princeton University Press Princeton

    1953

  16. [24]

    Olman, Lila Davachi, and Souheil Inati

    Cheryl A. Olman, Lila Davachi, and Souheil Inati. Distortion and Signal Loss in Medial Temporal Lobe. PLoS ONE , 4(12):e8160, December 2009. ISSN 1932-6203. doi:10.1371/journal.pone.0008160. URL https://dx.plos.org/10.1371/journal.pone.0008160. Publisher: Public Library of Sci...

    2009 doi

  17. [1598]

    URL https://www.physiology.org/doi/10.1152/jn.00338.2011

    doi:10.1152/jn.00338.2011. URL https://www.physiology.org/doi/10.1152/jn.00338.2011. Publisher: American Physiological Society

    2011

  18. [2017]

    URL https://www.pnas.org/doi/10.1073/pnas.1704877114

    doi:10.1073/pnas.1704877114. URL https://www.pnas.org/doi/10.1073/pnas.1704877114. Publisher: Proceedings of the National Academy of Sciences. 8 VIBE: Video-Input Brain Encoder

    doi

  19. [2025]

    arXiv:2503.20215 [cs]

    URL http://arxiv.org/abs/2503.20215. arXiv:2503.20215 [cs]

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.